Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017 ARM Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/Error.h" |
| 27 | #include "arm_compute/core/Helpers.h" |
| 28 | #include "arm_compute/core/ITensor.h" |
| 29 | #include "arm_compute/core/Types.h" |
| 30 | #include "arm_compute/core/Validate.h" |
| 31 | #include "arm_compute/core/Window.h" |
| 32 | |
| 33 | #include <arm_neon.h> |
| 34 | #include <cstddef> |
| 35 | #include <cstdint> |
| 36 | |
| 37 | using namespace arm_compute; |
| 38 | |
| 39 | namespace arm_compute |
| 40 | { |
| 41 | class Coordinates; |
| 42 | } // namespace arm_compute |
| 43 | |
| 44 | NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel() |
| 45 | : _input(nullptr), _output(nullptr), _result_offset(0), _result_mult_int(0), _result_shift(0) |
| 46 | { |
| 47 | } |
| 48 | |
| 49 | void NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::configure(const ITensor *input, ITensor *output, int result_offset, int result_mult_int, int result_shift) |
| 50 | { |
| 51 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32); |
| 52 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8); |
| 53 | |
| 54 | _input = input; |
| 55 | _output = output; |
| 56 | _result_offset = result_offset; |
| 57 | _result_mult_int = result_mult_int; |
| 58 | _result_shift = result_shift; |
| 59 | |
| 60 | constexpr unsigned int num_elems_processed_per_iteration = 16; |
| 61 | |
| 62 | // Configure kernel window |
| 63 | Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration)); |
| 64 | |
| 65 | AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration); |
| 66 | AccessWindowHorizontal output_result_access(output->info(), 0, num_elems_processed_per_iteration); |
| 67 | |
| 68 | update_window_and_padding(win, |
| 69 | input_access, |
| 70 | output_result_access); |
| 71 | |
| 72 | output_result_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->info()->tensor_shape())); |
| 73 | |
| 74 | INEKernel::configure(win); |
| 75 | } |
| 76 | |
| 77 | void NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::run(const Window &window, const ThreadInfo &info) |
| 78 | { |
| 79 | ARM_COMPUTE_UNUSED(info); |
| 80 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 81 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 82 | |
| 83 | const int32x4_t result_offset_s32 = vdupq_n_s32(_result_offset); |
| 84 | const int32x4_t result_shift_s32 = vdupq_n_s32(-_result_shift); |
| 85 | const int32x4_t zero_s32 = vdupq_n_s32(0); |
| 86 | |
| 87 | Iterator in(_input, window); |
| 88 | Iterator out(_output, window); |
| 89 | |
| 90 | execute_window_loop(window, [&](const Coordinates & id) |
| 91 | { |
| 92 | int32x4x4_t in_s32 = |
| 93 | { |
| 94 | { |
| 95 | vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + 0), |
| 96 | vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + 4), |
| 97 | vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + 8), |
| 98 | vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + 12) |
| 99 | } |
| 100 | }; |
| 101 | |
| 102 | // Add the offset terms to GEMM's result |
| 103 | in_s32.val[0] = vaddq_s32(in_s32.val[0], result_offset_s32); |
| 104 | in_s32.val[1] = vaddq_s32(in_s32.val[1], result_offset_s32); |
| 105 | in_s32.val[2] = vaddq_s32(in_s32.val[2], result_offset_s32); |
| 106 | in_s32.val[3] = vaddq_s32(in_s32.val[3], result_offset_s32); |
| 107 | |
| 108 | // Multiply by c_mult_int |
| 109 | in_s32.val[0] = vmulq_n_s32(in_s32.val[0], _result_mult_int); |
| 110 | in_s32.val[1] = vmulq_n_s32(in_s32.val[1], _result_mult_int); |
| 111 | in_s32.val[2] = vmulq_n_s32(in_s32.val[2], _result_mult_int); |
| 112 | in_s32.val[3] = vmulq_n_s32(in_s32.val[3], _result_mult_int); |
| 113 | |
| 114 | // Shift final result (negative value shift right) |
| 115 | in_s32.val[0] = vshlq_s32(in_s32.val[0], result_shift_s32); |
| 116 | in_s32.val[1] = vshlq_s32(in_s32.val[1], result_shift_s32); |
| 117 | in_s32.val[2] = vshlq_s32(in_s32.val[2], result_shift_s32); |
| 118 | in_s32.val[3] = vshlq_s32(in_s32.val[3], result_shift_s32); |
| 119 | |
| 120 | // Saturate negative values |
| 121 | in_s32.val[0] = vmaxq_s32(in_s32.val[0], zero_s32); |
| 122 | in_s32.val[1] = vmaxq_s32(in_s32.val[1], zero_s32); |
| 123 | in_s32.val[2] = vmaxq_s32(in_s32.val[2], zero_s32); |
| 124 | in_s32.val[3] = vmaxq_s32(in_s32.val[3], zero_s32); |
| 125 | |
| 126 | // Convert S32 to S16 |
| 127 | const int16x8x2_t in_s16 = |
| 128 | { |
| 129 | { |
| 130 | vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])), |
| 131 | vcombine_s16(vqmovn_s32(in_s32.val[2]), vqmovn_s32(in_s32.val[3])) |
| 132 | } |
| 133 | }; |
| 134 | |
| 135 | // Convert S16 to U8 |
| 136 | const uint8x16_t out_u8 = vcombine_u8(vqmovun_s16(in_s16.val[0]), vqmovun_s16(in_s16.val[1])); |
| 137 | |
| 138 | vst1q_u8(out.ptr(), out_u8); |
| 139 | }, |
| 140 | in, out); |
| 141 | } |